Platformatic

Putting audit logging on your AI gateway's critical request path is one of the most common production mistakes we see. Every audit write adds latency. Every audit failure risks your user-facing response. Here is how we solve it: blog.platformatic.dev/auditable-ai-g…

🚀 Autoscaling isn’t broken. It’s just too late. By the time HPA or KEDA reacts, users have already felt the spike. What if your infrastructure could predict traffic instead? New Banter episode with @lucamaraschi & @matteocollina 👇 streamyard.com/watch/jtj4vdxg…

📊 The difference with predictive scaling: ✅ 99.47% success rate ✅ 26 ms median latency ✅ No event loop saturation cliff (Compared to 95% / 154 ms for KEDA and 90% / 522 ms for HPA) Stop reacting. Start predicting. Explore White Paper ↓ arxiv.org/abs/2604.19705

Node.js makes this worse. CPU ≠ real load. The event loop can be saturated while the CPU looks “normal”. That’s where latency suddenly explodes. The cliff. Reactive scalers don’t see it coming.

🚀 Kubernetes autoscaling is always late. By the time HPA or KEDA reacts, your app is already overloaded. We built a different approach. Platformatic ICC predicts load trends and scales before the spike hits 👇 blog.platformatic.dev/ahead-of-time-…

This setup is simpler, and faster. Example We split Next.js image optimization into a separate worker, using the SAME codebase. ❌ No extra repo ❌ No duplication ❌ No noisy neighbour issues Better performance + easier scaling, without added complexity blog.platformatic.dev/run-medusa-kub…

Watt runs multiple Node.js apps as worker threads in a single process. That means: 🔨 No manual service wiring 🏎️ Fast internal communication 🔢 One runtime, multiple apps Frontend → backend calls stay inside the process, no external hops. Simpler and faster.

Running @medusajs in production gets complex fast. Frontend, backend, admin, image optimization, routing, Kubernetes… We simplified it with a Watt monorepo: 1️⃣ One repo 1️⃣ One build 1️⃣ One deployable platform Here’s how ↓ blog.platformatic.dev/run-medusa-kub…

Your AI integration works in the demo. In production, you're silently losing 15% of your request logs at peak, and your latency doubles every time a provider slows down. This is the operational wall every AI team hits. Here's how to get past it blog.platformatic.dev/auditable-ai-g…

We went stateful. More ops complexity, but that's exactly what Regina Coordinator handles, so devs don't have to. Here's the full architecture: blog.platformatic.dev/agents-in-prod…

Stateless works for simple request-response agents. But agents that write code, run it, install packages, manage a virtual filesystem? Each step builds on the last. Stateless means serializing the whole environment. That has a cost. blog.platformatic.dev/agents-in-prod…

Hot take: most "production-ready" AI agent frameworks aren't. Stateless architecture doesn't work for agents that build real in-process state. You need stateful session management. Change my mind. blog.platformatic.dev/agents-in-prod…

Idle sessions suspend to save resources. Crashes trigger auto-recovery. Rolling deployments back everything up first. AI agents that finally behave on Kubernetes. Full breakdown: blog.platformatic.dev/agents-in-prod…

Regina Coordinator solves this. It routes sessions, detects failures in 30 seconds, backs up agent state automatically, and restores conversations on healthy pods, completely transparent to the user.

Your Kubernetes pod dies mid-conversation. The agent was mid-task, packages installed, files written, and tools open. Session state is gone. That's a support ticket. That's a trust problem. It happens every deploy. blog.platformatic.dev/agents-in-prod…

Last week we shipped alongside Regina Coordinator, stateful session recovery for agents on Kubernetes. Reliable sessions + kernel-enforced security, inside Watt, on your own infra. blog.platformatic.dev/agents-in-prod…

Why it matters: agents run code generated at runtime. A naive setup starts the process, attaches enforcement, and hopes nothing happens in the gap. Any unprotected window is an attack surface. Kernel-level policy from instruction zero.

Most agent security setups have a flaw: there's a gap between process start and when enforcement kicks in. Even milliseconds matter. eBPF Sandbox eliminates that window. Policy is active before the first instruction runs. blog.platformatic.dev/agents-in-prod…